Allen Institute releases OlmoEarth v1.1, a satellite imagery model that cuts inference costs threefold

Allen Institute announced OlmoEarth v1.1, an efficiency-optimized version of its satellite imagery foundation model, on May 19, 2026. The release reduces per-inference compute costs by up to threefold relative to the original OlmoEarth (released November 2025) while maintaining equivalent performance on research benchmarks and real-world partner applications including mangrove change detection, forest loss classification, and country-scale crop mapping.

The efficiency gain stems from a redesigned tokenization strategy for Sentinel-2 satellite data. Where OlmoEarth v1 created separate tokens for each of three resolution bands (10m, 20m, 60m) per timestep per spatial patch, v1.1 combines all bands into a single token per timestep per patch, reducing token sequences threefold. Since transformer compute scales quadratically with sequence length, this architectural change produces material savings across pretraining, fine-tuning, and inference phases.

The naive approach to collapsing multi-resolution tokens caused a 10 percentage-point performance drop on m-eurosat kNN, a standard remote sensing benchmark. The team addressed this through modifications to the pretraining regimen—detailed in an accompanying technical report—that preserve the model's ability to learn cross-band relationships despite the reduced token dimensionality. The final models achieve v1-equivalent performance with one-third the computational footprint.

OlmoEarth v1.1 releases in three sizes: Base, Tiny, and Nano. All weights and training code are publicly available. The Allen Institute notes some task-specific regressions exist (documented in the technical report) but states that for applications where v1.1 performs comparably, users should observe significant speedup during model fine-tuning and deployment.